// Reference SMaLL-100 translator for iOS (onnxruntime-swift + swift-transformers). // // Swift Package deps: // .package(url: "https://github.com/microsoft/onnxruntime-swift-package-manager", from: "1.20.0") // .package(url: "https://github.com/huggingface/swift-transformers", from: "0.1.8") // // Bundle onnx/encoder_model.onnx, onnx/decoder_model_merged.onnx, tokenizer.json, // lang_tokens.json. This is a reference implementation of USAGE.md; adjust to your // onnxruntime-swift / swift-transformers versions. import Foundation import onnxruntime_objc import Tokenizers // from swift-transformers final class Small100Translator { private let env = try! ORTEnv(loggingLevel: .warning) private let encoder: ORTSession private let decoder: ORTSession private let tokenizer: Tokenizer // loaded from tokenizer.json private let langToId: [String: Int64] private let nLayers = 3, eos: Int64 = 2, decoderStart: Int64 = 2 init(encoderPath: String, decoderPath: String, tokenizer: Tokenizer, langToId: [String: Int64]) throws { let opts = try ORTSessionOptions() self.encoder = try ORTSession(env: env, modelPath: encoderPath, sessionOptions: opts) self.decoder = try ORTSession(env: env, modelPath: decoderPath, sessionOptions: opts) self.tokenizer = tokenizer self.langToId = langToId } func translate(_ text: String, to tgt: String, maxLen: Int = 128) throws -> String { // 1) ids = [tgt lang token] + tokenizer(text) (tokenizer appends =2) var ids: [Int64] = [langToId[tgt]!] ids += tokenizer.encode(text: text).map { Int64($0) } let mask = [Int64](repeating: 1, count: ids.count) // 2) encoder let idsT = try tensor(ids, [1, ids.count]) let maskT = try tensor(mask, [1, mask.count]) let encOut = try encoder.run(withInputs: ["input_ids": idsT, "attention_mask": maskT], outputNames: ["last_hidden_state"], runOptions: nil) let encHidden = encOut["last_hidden_state"]! // 3) greedy decode over merged decoder (KV cache) var out: [Int] = [] var cur = decoderStart, useCache = false var past = try emptyPast() for _ in 0..